Are setInputScale and setInputMean necessary ?
I'm currently testing this snippet and it works without them (I suppose these operation are already in the network model) while it does not work with them.. prediction is wrong (I suppose operation are doubled in this case)!

NOTE: considering discussion tensorflow/models#7727, only updated models work correctly: tensorflow/models#8057
NOTE: comparing to other SSD networks, these models expect normalized inputs ([-1, 1] input range):

@roccopietrini, Do you try the model from the model zoo or a custom one? Model ssd_mobilenet_v3_large_coco_2020_01_14 expects normalized input.

I tried with the model zoo model (ssd_mobilenet_v3_small_coco_2020_01_14), with just transfer learning for my custom class (so I just modified some parameters about the training in the pipeline.config, without any modification to the network). I generated the pbtxt file using this: https://github.com/opencv/opencv/blob/master/samples/dnn/tf_text_graph_ssd.py
I know the network expect normalized input, but I guess this is done by the network itself ? If I use setInputScale and setInputMea the prediction is wrong (the bounding box is much bigger), so I suspect the operations are doubled (because they already occur in the network..).

@roccopietrini, maybe something has been changed in the TensorFlow Object Detection API and published model does not reflect actual training code. Do you mind to open a separate issue so we won't miss it? Please add all the details - upload a newtork if possible, add origin image and expected results.

Hi, I don't think at this point that we have a bug in opencv .. so I'd better ask in stackoverflow because it's not fully clear to me weather the network include the normalization or not by default. I'll be back if I discover something .. :)

Hi,

using this exact code with installed:

numpy-1.18.5
opencv-python-4.2.0.34

I get error:

cv2.error: OpenCV(4.2.0) /io/opencv/modules/dnn/src/dnn.cpp:562: error: (-2:Unspecified error) Can't create layer "FeatureExtractor/MobilenetV3/Conv/hard_swish/add" of type "AddV2" in function 'getLayerInstance'

Any idea what did I miss?

@boris-savic, open .pbtxt and replace all "AddV2" to "Add"

@boris-savic, open .pbtxt and replace all "AddV2" to "Add"

Thanks. This did move things a bit forward but seems to get stuck on another issue:

[ERROR:0] global /io/opencv/modules/dnn/src/dnn.cpp (3066) getLayerShapesRecursively OPENCV/DNN: [Eltwise]:(FeatureExtractor/MobilenetV3/expanded_conv_3/squeeze_excite/mul): getMemoryShapes() throws exception. inputs=2 outputs=0/1
[ERROR:0] global /io/opencv/modules/dnn/src/dnn.cpp (3069) getLayerShapesRecursively     input[0] = [ 1 72 40 40 ]
[ERROR:0] global /io/opencv/modules/dnn/src/dnn.cpp (3069) getLayerShapesRecursively     input[1] = [ 1 72 1 1 ]
[ERROR:0] global /io/opencv/modules/dnn/src/dnn.cpp (3075) getLayerShapesRecursively Exception message: OpenCV(4.2.0) /io/opencv/modules/dnn/src/layers/eltwise_layer.cpp:209: error: (-215:Assertion failed) inputs[0][j] == inputs[i][j] in function 'getMemoryShapes'

Traceback (most recent call last):
  File "/home/boris/workspace/cv-dnn/detection.py", line 13, in <module>
    classes, confidences, boxes = net.detect(frame, confThreshold=0.5)
cv2.error: OpenCV(4.2.0) /io/opencv/modules/dnn/src/layers/eltwise_layer.cpp:209: error: (-215:Assertion failed) inputs[0][j] == inputs[i][j] in function 'getMemoryShapes'


Process finished with exit code 1

@boris-savic, please use the latest version of the script or see similar issue: #17058

Bonsoir, comment télécharger le modèle MobileNet-SSD v3 ?
Lorsque je clique sur poids ou configuration, rien ne se passe.